Application of Deep Learning Algorithms for BIPV Power Forecasting in Italy

In recent years, the efficiency of building integrated photovoltaic (BIPV) systems has significantly improved. However, BIPV remains unappealing to certain consumers because of its erratic and unpredictable nature. This paper examined a standard artificial neural network (ANN) and three advanced deep learning algorithms (DLAs): long short-term memory (LSTM), gated recurrent units (GRU), and convolutional neural networks (CNN), to forecast PV power in Rome, Italy. The models are trained and tested using a six-year dataset obtained from the PVGIS website. The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) are used to assess these models' performance. The findings demonstrate that GRU performs the best across all evaluation metrics. The GRU model achieved the lowest RMSE and MAE values of 28.40W and 12.53W on the training dataset and 15.31W and 6.62W on the testing dataset. In contrast, when training time is considered, ANN requires considerably less time than any other model. ANN's trainable parameters are approximately 1.95, 7.07, and 9.37 times less than those of CNN, GRU, and LSTM, respectively.